843b9a2a
by Simon Peyton Jones at 2026-05-22T12:00:39+01:00
Drastically reduce the use of the coercion optimiser
This MR addresess #26679 by calling the coercion optimiser much less often;
the coercion optimiser is expensive, and often ineffective, especially when
it is called repeatedly.
See Note [Coercion optimisation] in GHC.Core.Corecion.Opt.
Specifically
* Make coercion optimisation into its own pass, controlled by its own flag
`-fopt-coercion` like any other pass. The pass runs early in the pipeline.
* The Simplifier now contents itself with merely applying the current substitution
to a coercion, and checking for reflexivity.
* Kill off OptCoercionOpts, which is now unused.
Another important, but only loosely related, improvement
* In `GHC.Core.Subst.substIdBndr`, `GHC.Core.TyCo.Subst.substTyVarBndrUsing`,
and `GHC.Core.TyCo.Subst.substTyVarBndrUsing`, check for an empty substitution.
This keeps substitutions empty in the no-op case.
See Note [Keeping the substitution empty] in GHC.Core.TyCo.Subst
Compile time improves. Here are the compile-time allocation change over 1%:
Baseline
Test value Change
------------------------------------------------------------
CoOpt_Singletons(normal) 721,677,974 -7.9% GOOD
LargeRecord(normal) 1,268,094,410 -20.4% GOOD
T12545(normal) 772,150,124 -9.2%
T12707(normal) 777,720,918 -1.5% GOOD
T14766(normal) 918,228,179 -50.2% GOOD
T15703(normal) 318,141,541 +1.0% BAD
T18223(normal) 371,496,473 -7.6% GOOD
T1969(normal) 670,830,326 +2.2% BAD
T20261(normal) 573,444,953 -2.2%
T24984(normal) 87,836,660 -1.5%
T3064(normal) 171,636,592 -6.2% GOOD
T5030(normal) 148,301,062 -16.4% GOOD
T5321Fun(normal) 263,120,022 -3.3% GOOD
T8095(normal) 777,033,006 -71.3% GOOD
T9020(optasm) 222,159,574 -2.5% GOOD
T9630(normal) 873,422,584 -15.4% GOOD
T9872b(normal) 1,906,890,318 -2.9% GOOD
T9872b_defer(normal) 2,878,170,737 -2.1% GOOD
T9872d(normal) 356,978,798 -7.7% GOOD
TcPlugin_RewritePerf(normal) 2,132,138,060 -2.6% GOOD
geo. mean -2.7%
minimum -71.3%
maximum +2.2%
Metric Decrease:
CoOpt_Singletons
LargeRecord
T12707
T14766
T18223
T3064
T5030
T5321Fun
T8095
T9020
T9630
T9872b
T9872b_defer
T9872d
TcPlugin_RewritePerf
Metric Increase:
T15703
T1969
Serious attempt to avoid duplicating large coercions
Progress on not-inlining coercions
Add a type signature
More on exported coercions
More on coercion bindings
Allow in letrec; indeed make the invariants for letrec the same
as the invariants for top-level bindings.
ToDo: document this
Fix expression-equality for (Coercion co)
We were just saying "True" which is utterly wrong.
See Note [Equality for coercions] in GHC.Core.Map.Type
Wibbles on coercion bindings
Try a very cheap coercion optimiser
Comments -- almost all just about core invariants
Fuse optCoRefl and substCo
Maybe this will be better than either
Update Notes about core binding invariants
T26332 really should fail with -dlinear-core-lint
Try switching off the big optCoercion except in O2
Comments only
Try dropping all optimisation
The Simplifier just calls substCo, instead of optCoRefl
Flags for tracing coercion sizes
More optCoRefl
* Deal with submultiplicities
* In InitialPhase run even with empty subst
Update user manual
Care with the simple refl optimiser
Improve the gobbler
Fix warning
Add changelog